AI Hallucinations: Why Chatbots Make Things Up
Why AI chatbots state false things confidently, where it happens most, and seven practical ways to catch and reduce mistakes.
Ask a chatbot for the capital of France and it will be right. Ask it for a court case supporting your argument and it may invent one, complete with a convincing name, date and quotation. That second behaviour is what people call an AI hallucination, and understanding why it happens is the most useful thing a regular user can learn.
What a hallucination is
A hallucination is output that sounds right and is not. It can be:
- an invented fact, statistic or date;
- a quotation nobody said;
- a reference to an article, book or legal case that does not exist;
- a wrong calculation presented without hesitation;
- a false description of a document you uploaded;
- a claim that an action was completed when it was not.
The danger is the tone. Wrong answers arrive in the same confident prose as right ones.
Why it happens
A language model is trained to predict what text should come next. It learns the patterns of language from enormous amounts of writing, and those patterns include a great many facts. But it stores no list of facts to look up, and it has no built-in sense of the difference between "I know this" and "this sounds like the kind of thing that would be true".
Several things follow.
- It fills gaps with plausible material. Asked for a reference, it produces something shaped like a reference.
- It is trained to be helpful. An answer is usually rewarded over "I don't know".
- Its knowledge stops at a cut-off date. Anything later is unknown unless the tool searches the web.
- It goes along with your framing. If your question contains a false assumption, it often builds on it.
- Long conversations drift. Early details get lost or blurred.
Newer models hallucinate less than older ones, and tools that search the web or read your documents reduce the problem further. None has eliminated it.
Where the risk is highest
| Higher risk | Lower risk |
|---|---|
| Citations, case law, academic references | Explaining well-known concepts |
| Exact numbers, prices, dates | Rewriting or summarising text you supply |
| Quotations | Brainstorming ideas |
| Recent events | Drafting emails and outlines |
| Obscure people, places, products | Translating common languages |
| Medical, legal and financial specifics | Formatting and restructuring |
| Arithmetic without a calculator tool | Explaining code you paste in |
A useful rule: the more specific and checkable the claim, the more it needs checking.
Seven ways to reduce and catch mistakes
1. Turn on search or supply the source. Ask the tool to search the web, or upload the document and tell it to answer only from that. Grounded answers are far more reliable.
2. Ask for sources — then open them. A link is not proof. Models can cite real pages that do not say what they claim, or invent links entirely.
3. Give permission to say "I don't know". Add: "If you are not sure, say so instead of guessing." It helps more than you would expect.
4. Ask it to separate fact from inference. "Mark which statements come from the document and which are your own reasoning."
5. Ask again, differently. Start a new chat and rephrase. Answers that change between attempts are suspect.
6. Check the high-risk items yourself. Names, numbers, dates, quotations and references. Thirty seconds with a search engine or the original document.
7. Use a calculator for sums. Or ask the tool to run code for the calculation.
Prompts that help
- "Answer using only the attached report. Quote the passage that supports each point."
- "List your sources with links. If you cannot find a source for a claim, leave the claim out."
- "Before answering, tell me whether you are confident, and why."
- "What in your answer are you least sure about?"
More techniques are in how to write AI prompts.
When it matters most
- Health, law and money. Use AI to understand and to prepare questions, then confirm with a professional or an official source.
- Work you sign. Lawyers have been sanctioned for filing documents with invented cases. If your name is on it, you check it.
- News. Go to the original report. This site links its sources for that reason.
- Agents. When an AI acts instead of just answering, a mistake becomes an action. See what are AI agents.
Will it be fixed?
Hallucination is a consequence of how these models work, so it is being reduced, not removed. Three developments help: models that reason before answering, tools that retrieve documents and cite them, and systems that check their own output. The practical position for now is the one you would take with a knowledgeable colleague who sometimes bluffs: listen, then verify what matters.
Frequently asked questions
What is an AI hallucination?
An answer from an AI system that is false or invented but presented as fact.
Why does ChatGPT make things up?
Because language models generate text that is statistically plausible. They do not check facts against a database unless connected to search or documents.
How can I tell if an AI answer is wrong?
You often cannot from the tone. Check specific claims, such as names, numbers, quotations and references, against a reliable source.
Do newer models still hallucinate?
Yes, though less often. Search and document grounding reduce it further, but no model is free of errors.
This article was checked against the sources listed above on 12 November 2026. AI products change quickly: features, prices and availability can differ by plan and country, so confirm details with the company before relying on them. GetAIHow is independent and not affiliated with the companies it covers. Read our editorial policy and disclaimer.
